Van Guardian - Server

Van Guardian is a Ktor HTTP service that analyzes images of a camper-van interior and returns a concise safety assessment. It uses LLMs via Ollama to produce image descriptions and translations.

How it works

• Accepts an image as input.
• Produces a single JSON result describing whether a safety alert is present, a short reason, and a severity level (low / medium / high).
• (Optional) If an alert is present, the human-readable reason is translated to Spanish.

API contract

• /analyze: inputs an image and returns a JSON:

  • alert -> boolean: If LLM detects a safety issue
  • reason -> string: Human-readable explanation of the issue
  • severity -> enum (nullable): low, medium, high

Example request

curl -X POST --data-binary @photo.jpg http://localhost:8080/analyze -H "Content-Type: image/jpeg"

Example responses

No alert detected:

{
  "alert": false,
  "reason": "No suspicious activity detected"
}

Alert detected:

{
  "alert": true,
  "reason": "Smoke detected",
  "severity": "medium"
}

Ollama connection

Vanguardian requires access to an Ollama server to perform its analysis.

Setup instructions

1. Install Ollama: https://ollama.com/download

2. Start the Ollama server via CLI (for GUI instructions, go to 4):

   ollama serve

3. Load or pull the required models:

   ollama pull llava
   ollama pull gemma3

   By default, llava is used to describe the images and output the json, and gemma3 is used to translate the reason to spanish.

4. (Optional) You can also run it via GUI, by opening it and running any prompt using the required models, so the Ollama service is started and ready in the background

5. Ensure it is reachable:

   curl http://localhost:11434/api/tags

6. (Optional) If using a different host or port, set the service’s Ollama URL through configuration (OllamaUrl const).